Resilient multi-robot system with social learning for smart factories

Abstract

A system and methods for operating a multi-robot system (MRS) are disclosed. In some aspects, each robot of the MRS can: determine a local system regret state belief based on local evidence obtained by the robot itself and social evidence provided by other robots in a social community, determine a local system drift state belief based on the local system regret state belief, determine a next action based on the local system regret state belief and the local system drift state belief, and execute the next action. Local system regret state belief is generally an estimation of a system regret state for the MRS. Local system drift state belief is generally an estimate of a system drift state for the MRS.

What is claimed is: 
     
         1 . A method of operating a first robot in a multi-robot system, wherein the first robot and at least one additional robot of the multi-robot system form a social community in a network, the method comprising:
 determining, by the first robot, a local system regret state belief based on local evidence obtained by the first robot and social evidence provided by the at least one additional robot in the social community, wherein the local system regret state belief is an estimation of a system regret state for the multi-robot system;   determining, by the first robot, a local system drift state belief based on the local system regret state belief, wherein the local system drift state belief is an estimate of a system drift state for the multi-robot system;   generating, by the first robot, an opinion vector representing a level of trust to the at least one additional robot in the social community, wherein the opinion vector is calculated from the local system drift state belief;   receiving, by a remote computing device, the opinion vector;   generating, by the remote computing device, a trust matrix based at least in part on the opinion vector and at least another received opinion vector;   responsive to detecting, by the remote computing device and based at least in part on the trust matrix, a point failure resulting in product defects, initiating a corrective action within the multi-robot system, wherein the corrective action comprises at least one of forcing a robot calibration or scheduling maintenance;   determining, by the first robot, a next action relating to a manufacturing or fabrication task based on the local system regret state belief and the local system drift state belief; and   executing, by the first robot, the next action.   
     
     
         2 . The method of  claim 1 , further comprising:
 transmitting, by the first robot, the opinion vector to the remote computing device.   
     
     
         3 . The method of  claim 1 , wherein the local evidence comprises a least one of a local system regret state belief and/or a local system drift state belief determined by the first robot at a previous time step and measurements of a physical product obtained by the first robot. 
     
     
         4 . The method of  claim 1 , wherein the social evidence received from the at least one additional robot indicating a state of a product being fabricated or manipulated by the multi-robot system as determined by the at least one additional robot. 
     
     
         5 . The method of  claim 1 , wherein the social evidence comprises measurements of a physical product obtained by the at least one additional robot at a previous time step. 
     
     
         6 . The method of  claim 1 , wherein the next action is determined using a reinforcement learning model. 
     
     
         7 . The method of  claim 6 , wherein determining the next action is further based on a reward value provided by the remote computing device. 
     
     
         8 . The method of  claim 1 , wherein the local system regret state belief is determined using a Bayesian network. 
     
     
         9 . The method of  claim 1 , wherein the local system drift state belief is determined using a stochastic gradient descent (SGD) network with Huber loss. 
     
     
         10 . The method of  claim 1 , wherein the social network is a partially connected network that shares hypervertex set {R m,n } with the physical network, G phy . 
     
     
         11 . A control system for a first robot in a multi-robot system, wherein the first robot and at least one additional robot of the multi-robot system form a social community in a network, the control system comprising:
 a processor; and   memory having instructions stored thereon that, when executed by the processor, cause the control system to:
 determine a local system regret state belief based on local evidence obtained by the first robot and social evidence provided by the at least one additional robot in the social community, wherein the local system regret state belief is an estimation of a system regret state for the multi-robot system; 
 determine a local system drift state belief based on the local system regret state belief, wherein the local system drift state belief is an estimate of a system drift state for the multi-robot system; 
 generate an opinion vector representing a level of trust to the at least one additional robot in the social community, wherein the opinion vector is calculated from the local system drift state belief, and wherein the opinion vector is transmitted to a remote computing device configured to: (i) generate a trust matrix based at least in part on the opinion vector and at least another received opinion vector, and (ii) responsive to detecting, based at least in part on the trust matrix, a point failure resulting in product defects, initiate a corrective action within the multi-robot system, wherein the corrective action comprises at least one of forcing a robot calibration or scheduling maintenance; 
 determine a next action for the first robot relating to a manufacturing or fabrication task based on the local system regret state belief and the local system drift state belief; and 
 control the first robot to execute the next action. 
   
     
     
         12 . The control system of  claim 11 , wherein the local evidence comprises at least one of a local system regret state belief and/or a local system drift state belief determined by the first robot at a previous time step and measurements of a physical product obtained by the first robot. 
     
     
         13 . The control system of  claim 11 , wherein the social evidence received from the at least one additional robot indicating a state of a product being fabricated or manipulated by the multi-robot system as determined by the at least one additional robot. 
     
     
         14 . The control system of  claim 11 , wherein the social evidence comprises measurements of a physical product obtained by the at least one additional robot at a previous time step. 
     
     
         15 . The control system of  claim 11 , wherein the next action is determined using a reinforcement learning model, wherein determining the next action is further based on a reward value provided by a remote device. 
     
     
         16 . The control system of  claim 11 , wherein the local system regret state belief is determined using a Bayesian network. 
     
     
         17 . The control system of  claim 11 , wherein the local system drift state belief is determined using a stochastic gradient descent (SGD) network with Huber loss. 
     
     
         18 . A non-transitory computer readable medium having instructions stored thereon that, when executed by a processor, cause a device to:
 determine, for a first robot in a multi-robot system, a local system regret state belief, wherein the first robot and at least one additional robot of the multi-robot system form a social community in a network, wherein the local system regret state belief is determined based on local evidence obtained by the first robot and social evidence provided by the at least one additional robot in the social community, and wherein the local system regret state belief is an estimation of a system regret state for the multi-robot system;   determine, for the first robot, a local system drift state belief based on the local system regret state belief, wherein the local system drift state belief is an estimate of a system drift state for the multi-robot system;   generate, for the first robot, an opinion vector representing a level of trust to the at least one additional robot in the social community, wherein the opinion vector is calculated from the local system drift state belief, and wherein the opinion vector is transmitted to a remote computing device configured to: (i) generate a trust matrix based at least in part on the opinion vector and at least another received opinion vector, and (ii) responsive to detecting, based at least in part on the trust matrix, a point failure resulting in product defects, initiate a corrective action within the multi-robot system, wherein the corrective action comprises at least one of forcing a robot calibration or scheduling maintenance;   determine a next action for the first robot relating to a manufacturing or fabrication task based on the local system regret state belief and the local system drift state belief; and   control the first robot to execute the next action.